Articles
For a moment, the robot turns in a circle – but then something miraculous happens. It works out how it needs to adapt, using large swings of its thigh joints. After eight seconds, it can scuttle away.
The attack on the dog-like device was a dramatic demonstration of the new ‘omni-bodied robot brain’ from Californian robotic intelligence firm Skild AI. Unlike most robotic control systems, which are optimised for specific machines with set ‘morphologies’ (structures and mechanics), the new brain was trained on 100,000 different bodies, spending a virtual millennia learning how to move and control a variety of devices.
Today, most AI controllers learn specific locomotion strategies for individual robots. “This is like memorising the answer to a test – great for passing, but unhelpful for learning how to arrive at the answer,” a Skild blog post says. “When the AI faces a situation it has never seen before, like jammed motors, broken limbs or a new body, the memorised solution is useless and the AI doesn’t know how to fix it.”
A bird-brained idea?
Instead of memorising solutions for one body, Skild’s system finds and applies strategies from a “whole multiverse” of virtual robots, allowing it to adapt. Other demonstrations by the company include a quadruped having one or both legs disabled, or all of them lengthened. In all cases, the robot appears unstable for a few seconds before turning frantic motions into purposeful movements.
Taking training from 100,000 robots and applying it in one unit is a significant challenge, says Dimitrios Kanoulas, professor of robotics at University College London, who was not involved in the work. “It’s a little bit crazy,” he says. “It’s like having one brain for you and a bird.”
With no nerve endings, the robots cannot feel damage but learn from failures. “Then it understands that, ‘OK, now I’m a different robot, and I will use a different morphology from my learned scheme’,” Kanoulas says.
Mighty morphing
The Skild website reveals a broad spectrum of potential applications, from security and inspection quadrupeds for dangerous environments to humanoid robots for
in-home assistance and industrial work.
Such a system could be used “pretty much everywhere,” Kanoulas says, as “robots break very often”. Machines with the omni-bodied brain could respond
to damage like humans, quickly adjusting their actions to continue working before seeking repair if it gets worse. “What remains to see now is if this is scalable, robust and efficient,” he says. “Can I achieve a nice, robust, efficient locomotion, let’s say, with the same policy on 10 robots that have different morphologies?”
Control systems trained for specific robots might perform a bit better in some situations, but the Skild system could be well-suited to companies that are rapidly developing and deploying new hardware, controlling each new iteration regardless of whether it has six legs, wheels or any other number of new morphologies.
Want the best engineering stories delivered straight to your inbox? The Professional Engineering newsletter gives you vital updates on the most cutting-edge engineering and exciting new job opportunities. To sign up, click here.
Content published by Professional Engineering does not necessarily represent the views of the Institution of Mechanical Engineers.